WARP: On the Benefits of Weight Averaged Rewarded Policies

TL;DR

This paper introduces WARP, a novel method for aligning language models using weight averaging of policies at multiple stages, improving reward optimization while maintaining pre-trained knowledge.

Contribution

WARP proposes a new strategy that merges policies in weight space at three stages to better balance reward maximization and knowledge retention in RLHF.

Findings

WARP outperforms other open-source LLMs in reward and alignment quality.

Iterative weight averaging refines the reward-KL Pareto front effectively.

WARP improves policy performance at fixed KL divergence.

Abstract

Reinforcement learning from human feedback (RLHF) aligns large language models (LLMs) by encouraging their generations to have high rewards, using a reward model trained on human preferences. To prevent the forgetting of pre-trained knowledge, RLHF usually incorporates a KL regularization; this forces the policy to remain close to its supervised fine-tuned initialization, though it hinders the reward optimization. To tackle the trade-off between KL and reward, in this paper we introduce a novel alignment strategy named Weight Averaged Rewarded Policies (WARP). WARP merges policies in the weight space at three distinct stages. First, it uses the exponential moving average of the policy as a dynamic anchor in the KL regularization. Second, it applies spherical interpolation to merge independently fine-tuned policies into a new enhanced one. Third, it linearly interpolates between this merged model and the initialization, to recover features from pre-training. This procedure is then applied iteratively, with each iteration's final model used as an advanced initialization for the next, progressively refining the KL-reward Pareto front, achieving superior rewards at fixed KL. Experiments with GEMMA policies validate that WARP improves their quality and alignment, outperforming other open-source LLMs.